Machine Learning Identifies Two Gene Markers of Allergic Rhinitis

A computational analysis of public nasal-tissue expression data found that the genes TNFSF14 and CD83, when assessed together, separated samples from people with allergic rhinitis from healthy controls. The study used one discovery cohort and two external validation datasets, identifying a two-gene classifier that reached an area under the receiver operating characteristic curve, or AUC, of 0.968 in the training set, 1.000 in the smaller validation set GSE46171 and 0.758 in GSE40037. These results position the pair as a promising candidate panel for further development, although the authors caution that the analysis is retrospective, based on modest numbers of stored samples, and does not establish that the genes cause disease or are ready for clinical diagnosis.

The research team began with a microarray discovery cohort, GSE75011, containing 15 samples from people with allergic rhinitis and 25 healthy controls. Using a deliberately inclusive screening threshold of p below 0.05 and an absolute log2 fold change above 0.585, the authors identified transcripts that differed between cases and controls. They intersected these differentially expressed genes with a curated set of 1,542 cell-adhesion-related genes from the Molecular Signatures Database, producing 38 differentially expressed adhesion-related genes. Gene Ontology and Kyoto Encyclopedia of Genes and Genomes analyses associated this group with leukocyte adhesion, cell activation, cytokine-receptor binding, focal adhesion and immunoglobulin-superfamily cell-adhesion-molecule signaling. The researchers emphasize that the strategy was designed to capture immune-regulatory programs involved in cell-cell interaction rather than to restrict attention to structural adhesion proteins alone.

Protein-interaction mapping provided a second layer of prioritization. The investigators queried the STRING database, retained interactions with confidence scores of at least 0.15, visualized the network in Cytoscape and used the MCODE plug-in to identify densely connected modules. MCODE detected two modules, and the highest-scoring module contained 13 candidate hub genes. All 13 showed statistically significant expression differences between allergic-rhinitis and control samples in the discovery cohort. CX3CR1 and CADM1 were elevated in allergic rhinitis, whereas NR4A3, PDCD1, MAP3K8, TNFSF14, ZC3H12A, DUSP1, NLRP3, SMAD7, TBX21, CD83 and ICOSLG were reduced. Functional annotation linked the hub candidates to CD4-positive alpha-beta T-cell differentiation, T-cell activation and leukocyte cell-cell adhesion, extending the analysis from adhesion-associated expression toward the immune circuitry in which those signals sit.

The 13 hub candidates were then screened in parallel by least absolute shrinkage and selection operator regression, random forest and support vector machine recursive feature elimination. LASSO retained five genes at its cross-validated penalty, random forest ranked six above a MeanDecreaseGini threshold of 1, and SVM-RFE achieved its lowest classification error with three features. The overlap across all three algorithms consisted of TNFSF14 and CD83. In the discovery cohort, TNFSF14 alone produced an AUC of 0.931 and CD83 an AUC of 0.845; the combined model reached 0.968. The authors also incorporated the two markers into a nomogram for estimating sample-level probability, examined calibration between predicted and observed status and used decision-curve analysis to assess performance across probability thresholds.

External testing produced variable but encouraging results. The two-gene classifier achieved an AUC of 1.000 in GSE46171, which included six allergic-rhinitis samples and three controls, but the authors explicitly warn that such a small cohort warrants cautious interpretation. Its AUC was 0.758 in GSE40037, a dataset comprising 10 cases and 12 controls, indicating that discrimination persisted but was less than in the discovery set. At the individual-marker level, CD83 retained moderate diagnostic value in both validation cohorts, whereas TNFSF14 performance varied. Both genes tended to be lower in allergic-rhinitis samples overall, yet their group differences were not identical across datasets: TNFSF14 remained significantly reduced in GSE46171 but not GSE40037, while CD83 was significantly reduced in GSE40037 but not GSE46171. The combined panel therefore appeared more reproducible than either gene alone in this analysis.

Both markers are embedded in immune pathways relevant to allergic inflammation. TNFSF14 encodes a TNF-superfamily ligand that binds herpesvirus entry mediator and lymphotoxin-beta receptors to stimulate lymphocyte proliferation and modulate immune responses. Its reduced expression in allergic rhinitis is not straightforward to interpret because the gene has been associated with immune activation in other settings; the authors say altered immune-cell composition, disease stage or compensatory responses could account for the observation. CD83 is associated with dendritic-cell maturation and regulation, and both membrane-bound and soluble forms can influence T- and B-cell responses. Lower bulk-tissue expression may likewise reflect differences in dendritic-cell abundance or composition rather than reduced regulation within individual cells. The researchers therefore interpret both signals as immune-associated transcriptional features, not direct evidence of impaired molecular pathways or causal contributions to disease.

Immune deconvolution suggested that the expression differences sat within a remodeled T-cell environment. Single-sample gene set enrichment analysis showed lower activated dendritic-cell and T follicular-helper signatures in allergic-rhinitis samples, alongside a moderate increase in a general T-helper signature and a reduction in T-cell co-stimulation scores. CIBERSORT produced a complementary view, with lower inferred levels of resting memory CD4 T cells and higher levels of naive CD4 T cells in cases. Pearson correlations showed that TNFSF14 and CD83 had distinct but overlapping relationships with multiple immune-cell estimates. These computational results are consistent with a changed cellular and functional context in the nasal samples, but they do not establish which immune populations cause the expression shifts. Bulk transcriptomics can make changes in cell abundance or composition resemble altered regulation within particular cell types.

Expression-based clustering further divided the allergic-rhinitis samples in the training cohort into two exploratory groups. Non-negative matrix factorization indicated a two-group structure, and principal component analysis placed the groups in different transcriptomic regions. TNFSF14 expression was higher in Subtype 2, whereas CD83 was higher in Subtype 1. The groups also differed in single-sample estimates of plasmacytoid dendritic cells, helper T cells and tumor-infiltrating lymphocyte-related signals, as well as in antigen-presenting-cell co-stimulation, checkpoint activity, cytolytic activity and T-cell co-stimulation. MCP-counter indicated broader differences in inferred immune-cell composition. Genes distinguishing the groups were enriched for depyrimidination, ketone catabolism and small-molecule catabolic processes, pointing to possible metabolic differences alongside immune variation. Because clustering was based on only the samples in the discovery cohort, the authors describe these groups as exploratory molecular patterns rather than definitive clinical subtypes.

The study also mapped possible regulatory influences, but these results remain hypothesis-generating. NetworkAnalyst predictions connected both markers to candidate transcription factors, including members of the NF-kappa-B, STAT and IRF families, while experimentally supported miRNA interactions from miRTarBase were combined with lncRNA-miRNA relationships to construct a putative competing endogenous RNA network. TNFSF14 occupied a more central position and was linked to more predicted non-coding RNA regulators than CD83. Enrichment against DSigDB associated the expression signatures with compounds including puromycin, menadione, sanguinarine and betamethasone valerate, but database similarity does not establish that any compound treats allergic rhinitis or alters these markers in patients. The analysis also lacks protein measurements, functional experiments, detailed severity and treatment information, control for comorbid asthma or conjunctivitis, and broad representation across regions and populations.

Taken together, the findings offer a two-gene computational signature and a possible framework for understanding molecular variation within allergic rhinitis. TNFSF14 and CD83 warrant experimental validation in independent nasal-tissue or blood samples, assessment of whether their RNA levels correspond to protein abundance, and testing of whether manipulating either gene changes dendritic-cell or T-cell behavior. Prospective studies with larger and clinically annotated cohorts would also be needed to test whether the panel retains accuracy in people receiving different treatments, those with different symptom burdens and those with coexisting allergic disease. For now, the strongest conclusion is that a cross-algorithm analysis of adhesion-associated immune expression consistently selected TNFSF14 and CD83, with functional and clinical relevance still awaiting direct evaluation.

Subject of Research: Transcriptomic and immune-signature analysis of candidate diagnostic biomarkers in allergic rhinitis

Article Title: Integrative Network Analysis and Machine Learning Identify Cell Adhesion‐Related Biomarkers and Immune Heterogeneity in Allergic Rhinitis

Article References: He, G., & Zhang, L. (2026). Integrative Network Analysis and Machine Learning Identify Cell Adhesion‐Related Biomarkers and Immune Heterogeneity in Allergic Rhinitis. Immunity, Inflammation and Disease, 14(10), Article e70562. https://doi.org/10.1002/iid3.70562

Image Credits: AI Generated

DOI: 10.1002/iid3.70562

Keywords: allergic rhinitis, TNFSF14, CD83, diagnostic biomarkers, machine learning, transcriptomics, immune heterogeneity, cell adhesion, T-cell profiling, public gene-expression datasets, Integrative, Network